8.3.1 Scenario 1: Agents for the Compensation of Sensor Contamination

In industrial production processes that involve high dust or vapor loads, the reliability of applied sensors (e.g., photo sensors) is a critical issue. Because a contamination of the sensors may arise from the surrounding production process and strongly affect the functionality of the sensor, the control software may obtain incorrect sensor values. Different strategies for the automation software to appropriately react to such sensor failures exist (e.g., shutting down the production system for maintenance or forcing the controlled device into a stable state). However, these strategies would result in downtimes for the production systems and therefore financial losses.

A solution to avoid downtimes in case of sensor failures is possible through an implementation for the automation software that integrates redundant, virtually calculated soft-sensor values to detect and compensate for failures of the real sensors. Although this may allow the continuation of a production process until the next maintenance scheduled, depending on the accuracy of the soft-sensor models the precisions of the soft-sensor values may be lower than the ones of the real sensors. Using the inherent ability to encapsulate knowledge, as well as mechanisms, to evaluate this knowledge inside software entities, the agent-oriented paradigm can be considered an appropriate means of implementing such an approach for the control software of production systems.

Due to its knowledge base, which is able to take various levels of contamination or processes of deterioration into account, agents are able to detect whether a change in a sensor signal is related to the material detected (a correct functioning of the sensor) or is due to increasing contamination or deterioration and, based on the soft sensors in their knowledge base, are able to compensate for the signal fluctuations. If the level of contamination is too high, or the lifetime of a sensor has been reached to a level of 90%, the agent could, for example, issue a warning for the sensor. Thus, it would be possible to reach a higher precision during processing and it would become easier to coordinate the maintenance and cleaning of the system.

8.3.2 Scenario 2: Agents for the Plug-and-Produce Configuration of Production Units

A second scenario may be described by the production of an individually planned kitchen, where adaptations of the production system become necessary during production. Therefore, flexible, adaptive, self-organizing software entities are required. By applying an agent-oriented paradigm to develop the automation software for the production systems, intelligent and autonomous objects, which take care of organizational tasks via a common network and fulfill the mentioned requirements, can be developed. The developed agents therefore could be aware of their functionality and boundaries, their position, their consumption of resources, rising operation costs, and their configuration alternatives.

Like those provided by the agent-oriented paradigm, the agents of production entities could offer their processing capabilities using a common interface entity (e.g., a whiteboard), identify other processing capabilities available, and negotiate regarding the execution of production steps. When the plant receives a request for production (e.g., the production of a customer-tailored kitchen), each production entity checks whether necessary production operations can be fulfilled by their capabilities. The relevant production entities respond and negotiate the general requirements under which they can set up the required production process. The result of the negotiations is the optimum configuration for fulfilling the production task, in which the production entities can carry through the negotiations completely via a planning agent. The fundamental planning information such as cost, availability, and production time is provided for the negotiations.

If, for instance, scratches are detected on the surface of a produced wooden plate, it has to be replaced. The different parts of a production system involved, such as their agents, receive the relevant product quality requirements for optimizing their own behavior. Using the inherent features of the agent-oriented approach, software agents could analyze these requirements by using their own knowledge base to evaluate possible reconfiguration to their part of the production system that ensures the required product quality. Based on the evaluations, the agents could reconfigure and adapt the parameters of the plant sections that they control (e.g., extruding temperature, contact pressure for coating, process speed).

As another part of this scenario, breakdowns of plant machinery could be compensated for by the agent-oriented implementation. If, for example, the agent of a high-performance saw inside a production system reports a failure, an enquiry for a saw with equivalent performance could be made inside the agent system. If no appropriate high-performance saw is available, the agents of two saws with lower performance could dynamically form a coalition and report back to collaboratively offer the service previously fulfilled by the high-performance saw. Consequently, using this behavior of the agents, the overall production system autonomously substitutes the high-performance saw with two lower-performance saws and continues to produce the furniture parts without interruption and without negative effects on efficiency and quality.

8.3.3 The Assessment of Potentials and Challenges in Industry

For the guided expert interviews, a questionnaire was developed based on the results and experiences from the previous research project AVE to introduce agent technology and possible applications to industrial practitioners in the field of production automation software development. The questionnaire required them to rate the importance of seven key features that condense the investigated methodologies and results from the project AVE (Wannagat and Vogel-Heuser, 2008) in a descriptive manner. The features can be divided into features that improve the performance of an industrial production system by applying software agents and features that are provided during the development of automation software. Concepts for increasing the dependability of a production system were one major result of the project AVE, previous to the investigations on applications and tool development inside KREAagentuse. The aspect product quality, which is considered an important issue in general, was integrated to have a benchmark to compare the relevance of plant dependability and energy efficiency. Subsuming, regarding the operation of a production system, the following features had to be rated:

(F1) The enhancement of product quality

(F2) The increase of a production plant’s dependability

(F3) The enhancement of a production plant’s energy efficiency

Models and concepts for soft-sensor redundancy, which were previously developed, support the implementation of diagnosis algorithms to calculate the impact of sensor failures on the automation software and also consider process requirements and their impact on the implementation for plant modules. Support for structuring automation software, as provided by related approaches for AOSE, such as Gaia (Lüder and Peschke, 2007), was considered inside the project AVE as well. Furthermore, concepts for the flexible allocation of process functions to execute system components similar to approaches for SOAs were developed. Concluding this, the features provided by the agent-based development of automation software that partly resulted from the project AVE as well had to be rated by the industrial practitioners and comprised the following:

(F4) Diagnosis of failure impacts inside the automation software

(F5) Flexibility concepts (e.g., SOAs)

(F6) Support for the appropriate structuring of automation software

(F7) The impact of process requirements on plant modules

In the following section, the results of the conducted interviews are presented (i.e., the relevance of the aforementioned seven features as rated by the industrial experts).

8.3.4 Results of the Survey in Machine and Plant Automation

The experts rated the monitoring and increase in product quality as most important in comparison to the increase of a plant’s dependability and energy efficiency (see Figure 8.1). Its importance is scarcely higher than the increased plant dependability. Increasing the energy efficiency of a production plant is rated as being important, but not very important.

In order to fulfill these requirements, the significance of model contents and tool requirements has been evaluated (cf. Figure 8.2). Here, the experts found the online-debugging function for tracing failure impacts most important, closely followed by the application of the flexibility concepts for distributed systems and the support of the modular structuring of the automation software. They are followed by a simple description of the constraints and limitations implied on a module’s operations and variables, which were rated as important to very important by a majority of the interviewees. The experts quite disagreed as far as the compensation of runtime errors by given function blocks respective to the models concerned. Continuing processes, which have to expect bigger problems when dropping out, are considered especially important. For other processes, compensation during runtime only plays an inferior special role, because it is generally decided in the development phase how it is to react to sensor breakdowns. However, the results of the expert evaluation confirm the customer relevance of agent-based systems and approaches in machine and plant engineering, as shown for the application scenarios and in the project KREAagentuse (Frank et al., 2011).

Along with the significance of the requirements, the challenges in turning new approaches into reality were part of the scope of the survey. In this context, the interviewees explained that the current systems require their individual diagnostic evaluation and producer-specific user guidance. For the implementation of an integrated debugging concept in the control systems of tomorrow, they consider standardization across all modules of all producers a must. There were complaints that the current web-diagnosis of the equipment is hardly usable, because each of the machines/each producer follows different concepts. Concerning the concepts of distributed systems and variable possibilities for supporting the coupling of single modules of the control software, deficits of today’s methods for the structuring of the program codes were remarked. According to the interviewees’ opinions, this was regarded as a future challenge, because new technologies and possibilities could only be used effectively, if significant progress was made in this direction. The modularity, which used to be possible with individual entities (in hardware), needs to become applicable to numerous individual application modules on hardware platforms in the future. In order to support the modular structuring of control software, integrated software-design tools, such as tools of UML or SysML, were requested. New tools should provide integrated support along the entire life cycle of the software. For instance, it needs to be possible to generate code from a UML/SysML-diagram, which can then be debugged and monitored online in the UML/SysML-diagram (Witsch et al., 2010).

Although not all presented features of software agents in manufacturing automation were rated highly relevant by the industrial experts, the interview results generally underpin the approaches and results from the project AVE. In the following section, further concepts for software agents in manufacturing automation are presented, which address the identified challenges and potentials by elaborating on and extending the results from AVE.

8.4.1 An Enhanced Concept for Soft-Sensor Estimation Using KDE

An approach presented in previous works (Schütz et al., 2013; Wannagat and Vogel-Heuser, 2008) proposed the description of physical dependencies between the different sensors and actuators of a manufacturing system. During a plant’s runtime, software agents can calculate soft-sensor values (i.e., using time dependencies between signals of two light barriers on a transportation belt) based on the redundancy model (in that case, this would describe the physical equation of movement between both light barriers). The software agent compares the real sensor value to the expected calculated soft sensor. After multiple observations, the deviation of the sensor signals result in a probability density estimation that is used to describe the accuracy of the soft sensor. An enhanced diagnosis and compensation of sensor failures (cf. feature F2), as described in Scenario 1 of the conducted interviews (see Section 8.3), can be realized by that approach and evaluated (Schütz et al., 2013; Wannagat and Vogel-Heuser, 2008).

The Gaussian distribution is proposed as the mathematical density estimation (Schütz et al., 2013; Wannagat and Vogel-Heuser, 2008). A failure is detected if the real sensor value is above a predefined threshold. In case of a value above this threshold, a failure of the real sensor is detected and the virtual sensor is activated in order to replace the real sensor. Especially for real-time applications, the assumption of a Gaussian normal distribution is beneficial due to the easy calculation that does not require high-performance computation. However, one major drawback of Gaussian distributions is the effect from the influences of erroneous data on the (one, two, and three) standard deviation, resulting in a changed shape (broadening) of the corresponding bell curve. Furthermore, due to the required symmetric characteristic of a Gaussian distribution, erroneous data are changing both sides of the Gaussian distribution, and thus the accuracy of the sensor value estimation strongly decreases. This increases the risk of false alarms (e.g., declaring a real sensor to be faulty/working even if this sensor is working correctly or is faulty).

As a solution to overcome these disadvantages, the use of kernel density estimation (KDE) is proposed. KDEs are a nonparametric way to estimate the density of random variables and was introduced by Rosenblatt (Rosenblatt, 1956) in 1956 and further developed by Parzen (1962). A KDE is a continuity function and consists of a number of random variables, the bandwidth, and a kernel function. The kernel function itself is a symmetric function. The most common kernel functions are the Gaussian, Epanechnikov, Biweight, Triweight, and Cosine functions. The criterion to select a kernel could be the minimization of the mean squared error (MSE) or the mean integrated squared error (MISE). One of the major benefits of using KDEs to estimate the density of process variables is that the KDE is not restricted to being symmetrical. Due to these characteristics, the distribution of process values such as sensor or soft-sensor values can be estimated more exactly than would be the case when using the Gaussian distribution. This also enables a more precise diagnosis of failure impacts inside the automation software. Figure 8.3 shows an application example that evaluates the estimation of the distribution (quality) of sensor values by using the Gaussian distribution and the KDE. The application example uses process data (sensor values) gathered from the crane module that lifts and transports work pieces inside a small laboratory manufacturing system (see Figure 8.3, right), the pick and place unit (PPU, cf. (Legat et al., 2013)). The crane consists of positioning sensors to detect if it is in its upper or lower position. It is able to turn clockwise and counterclockwise to transport the work pieces inside the PPU. Consequently, three digital positioning sensors are mounted on the bottom plate of the crane to indicate its current angle. The benefit of the KDE approach can be demonstrated for the timing behavior of the rotation of the crane. From a series of measurements focused on the time that the crane needs to execute a turn of 90 degrees, it was evaluated that it needs less time to turn clockwise than to turn counterclockwise.

In Figure 8.3 (left), the KDE (solid line) for the elapsed time of turning clockwise and turning counterclockwise in comparison to the Gaussian distribution (dashed line) is compared. This analysis proves that the KDE, like the Gaussian distribution, can also be used for accurately estimating the distribution. In comparison, the Gaussian distribution assumes the mean value of the estimation at 2.65 s, while the KDE points out that no data point has been observed in the area. Furthermore, the KDE has been formed from two different concentrations of data points. Additional analysis shows that one concentration represents the clockwise data points and the other the counterclockwise data points. The analysis of this application example indicates that the KDE approach for some systems may enhance the soft-sensor concepts developed in previous works. Although the KDE leads to higher complexity concerning mathematical calculations of standard deviations, confidence levels, and probability distributions, given the experiences gained from the first evaluations (e.g., considering the application for the PPU), the mentioned advantages of KDE have compensated for this drawback. Hence, the KDE approach for soft-sensor value estimation can be considered a promising approach to enhancing a manufacturing system’s dependability (cf. feature F2) and provide an enhanced failure (impact) diagnosis for manufacturing systems (cf. feature F4).

8.4.2 Model-Based Development of Agents and Soft Sensors

The model-based method developed in KREAagentuse extends the meta-model, diagrams, and notations of both the Unified Modeling Language (UML) and the SysML. The language profile SysML was developed to support the physical systems’ design (Espinoza et al., 2009) and has been widely applied inside different model-driven engineering approaches in the domain of production automation (Thramboulidis, 2010). For handling an agent system’s complexity and offering support for software structuring (cf. feature F6), the approach comprises an architectural model that is separated into four different views: the technical process (TP), the technical system (TS), the automation system (AS), and the function layer (FL). Each represents one major aspect of the agent system (cf. Figure 8.4). These views do not necessarily correspond to different engineering disciplines—for example, those presented by Thramboulidis (2010). They are instead concerned with different aspects of the agent system that control a production system.

A process agent’s functionality (i.e., the TP) is described as the behavior of the process agent in the form of an activity diagram (AD), containing sequenced and/or parallel activities that invoke the functions allocated on the FL by the resource agents. The machine functions of a manufacturing system for treating material or assembling products are allocated by software agents and described as functions on the FL of the proposed layered architecture. Following the meta-model of SysML, the model concept “requirement” can be used to define the requirements regarding product quality (cf. feature F1) and operational constraints that process functions must comply with (cf. feature F7). The functions themselves are declared inside block definition diagrams (BDDs) at the FL and can be implemented either by using the languages of IEC 61131–3 or by using the behavior diagrams of UML (e.g., Ads) or SysML parametric diagrams (PDs). This also applies to the implementation of the resource agents and the blocks of the TS, as well as to the implementation of blocks of sensors and actuators.

By defining that the block of an agent allocates a function, it is indicated that this agent composes its module’s operations to implement this function. Inside the TS, the technical structure and behavior of a manufacturing plant is modeled by hierarchically composing a plant’s components with their interconnections. The top level of this hierarchy is formed by the modules controlled by the software agents. Inside BDDs, these modules are composed from sensor and actuator modules on the field level. The automation system view comprises the real sensors and actuators of a manufacturing system as well as SysML blocks, which implement soft sensors used to detect and compensate for sensor failures in order to increase a plant’s dependability (cf. feature F2). The information necessary to implement these soft sensors is described inside two knowledge models, the redundancy model and the tolerance model, which are both illustrated using PDs. Whereas the redundancy model describes the analytical dependencies between soft and real sensors, the tolerance model enables the modeling of the impact of sensor failures on a module’s functions (cf. feature F4) (Schütz et al., 2013; Wannagat and Vogel-Heuser, 2008), as well as the monitoring of a product’s quality (cf. feature F1) as it strongly relates to the correct execution of the different process functions.

As part of the evaluation experiments, an expert interview with two industrial experts from machine and plant manufacturing, as well as two industrial experts from PLC and tool vendors, was conducted. For the interview, the developed approach was presented to the experts using the application example from Frank et al. (2011), followed by a discussion on the approach’s applicability. The experts gave positive feedback about the model-based approach in general, but criticized that it lacks support for the identification of an appropriate module and agent size (i.e., at which level of the software module hierarchy agents should be applied to control corresponding plant components).

8.4.3 The Generation of Control Strategies from Agent Models

In order to better support the flexibility of manufacturing systems (cf. feature F5) through the design of the automation software and thus enable implementing all the degrees of freedom an agent can apply to controlling a manufacturing system, an extension of the previously presented approach to develop the space of action of a module and the manufacturing system in a consolidated way was developed (Schütz et al., 2012) (cf. Figure 8.5). This concept aims at finding production strategies that can be optimized regarding different criteria (e.g., energy consumption) (cf. feature F3).

In the proposed extended architecture, a module is able to perform various operations. To enable the consolidated description of an agent’s complete space of action for controlling a module, it is required to model what a module is able to do instead of the traditional way of modeling or implementing what a module has to do. Therefore, restrictions on an operation’s execution are annotated within the SysML model by preconditions (a set of states allowing an operation’s execution) and the so-called effects that describe how the operation behaves (the transformation rule about how the products’ and system’s state are manipulated by an operation).

States for defining restrictions refer to possible measurements of a module’s sensors, (i.e., the data type of the sensor variable), as well as on internal control variables of the module’s agent. The dependencies between two agents’ function execution is also required to be constrained when composing a plant by modules. The precondition and effects of functions can be restricted specifically to a plant’s characteristics. Furthermore, composition constraints on functions (e.g., their parallel execution) can be defined. This lets valid processes automatically be executed by the MAS. This provides a strict modular development, while considering plant-specific characteristics (F6, F7).

Ontology-based formal model semantics were defined to facilitate the automatic processing (cf. Figure 8.6) of the model for automatic synthesis of an agent’s functions (Legat et al., 2014). Because the functions themselves are limited by restrictions as well, the overall process of a plant can be synthesized accordingly. As proposed in (Legat and Vogel-Heuser, 2013), this knowledge model can be applied to flexibly control manufacturing systems. This enables reasoning about possible compensation strategies in case of complete module failures (cf. feature F2), synthesizing production processes for changing product requirements (cf. feature F5), and optimizing a production process regarding the energy consumption of the plant to increase the energy efficiency (cf. feature F3). First, experiments for applying the approach have proven that although the time needed to synthesize a reconfiguration strategy or a production process previously not considered highly depends on the size and complexity of the considered manufacturing system, the approach can successfully be applied, and for small-scale manufacturing systems, a computing time within the range of the necessary 5-10 ms is needed (Legat et al., 2014).